Clover-35B-A3B-Preview, a diffusion language model from Diffusion Intelligence, can rapidly write code, generate web pages, call tools, and execute tasks like an agent, delivering noticeably faster performance than competing AR models in benchmarks.

The model runs on a Mixture of Experts architecture with 35 billion parameters, activating roughly 3 billion parameters per token, and uses a block diffusion approach with bidirectional attention to enable parallel generation while managing historical context.

Diffusion-based language modeling enables parallel updates across positions and iterative refinement through diffusion denoising, with Clover refining content via block-level diffusion, historical memory integration, and a gated delta network to balance new output with past context.

In agent-style tasks, Clover handles multiple tasks continuously via WebMCP, completing six tasks in sequence—from invoice issuance to information retrieval—and outperforming Gemini 3.6 Flash in speed and marginally higher success rates.

Diffusion Intelligence, founded by Kong Lingpeng and collaborators and building on Dream 7B, aims to expand Clover into multi-modal and end-side deployments with ongoing model scaling and deployment optimizations.

Clover can generate interactive web interfaces for a personal agent in about ten seconds and achieves strong results on code benchmarks such as SWE-bench Verified (73.6) and Terminal-Bench 2.1 (51.7).